There is a version of this article that opens with: "Here are 20 prompts to supercharge your qualitative analysis with ChatGPT." This is not that article.
ChatGPT can be genuinely useful for specific, bounded tasks in a qualitative workflow — and this guide provides prompts for those tasks. But general-purpose AI tools are structurally poorly suited to rigorous qualitative data analysis, and understanding why before you reach for a prompt library is worth your time. Researchers who understand the limitations use these tools effectively. Those who do not tend to produce analysis that looks thorough but is not, which is arguably worse than doing the work manually.
The prompts below are organised by task. Each section includes notes on what to watch for. Read the limitations section before using the prompts on anything that matters.
Why the "just prompt it" approach doesn't work for qualitative analysis
The problem is not that ChatGPT is insufficiently clever. The problem is architectural — and it is worth understanding at a mechanical level before deciding how to use it.
ChatGPT is a next-token prediction model. It generates text that is statistically likely given its training data and your prompt. It does not "read" a transcript in the way a researcher does: building an interpretive framework, tracking contradictions, noticing what was said hesitantly versus confidently, relating individual statements to a developing understanding of the whole. What it does is produce output that looks like qualitative analysis because it has been trained on large quantities of qualitative research writing. The distinction matters for how you interpret its output. How LLMs actually analyse data covers the mechanics in detail.
A second architectural constraint is the context window. For GPT-4o, this is around 128,000 tokens — roughly 90,000-100,000 words. A single interview transcript typically runs to 5,000-8,000 words; a corpus of 20 interviews would be 100,000-160,000 words, at the edge of or beyond what fits in a single context. Beyond 5-8 full transcripts, you have to chunk the data. Chunking means the model is not maintaining a coherent view of your full dataset as it processes each piece — which is precisely what cross-corpus thematic analysis requires.
The third constraint is statefulness. There is no persistent project. Each session starts fresh. The "codes" from Monday's session are not automatically carried into Wednesday's unless you paste them back in, which means consistency across a large coding project depends entirely on your own discipline in managing the codebook manually.
The practical implication is that ChatGPT works well for tasks that are bounded (one transcript, one question), where consistency across the corpus is not required, and where you can verify the output yourself. It works poorly for anything that requires coherent analysis across a full dataset. AI text analysis tools: why general-purpose tools aren't built for serious qualitative work develops this argument more fully, and is worth reading before committing to a ChatGPT-based analysis workflow for a substantial project.
Prompts that work: transcript summarisation
Transcript summarisation is the task ChatGPT handles most reliably, and the one where the architectural limitations matter least. The output does not need to be consistent across transcripts in the same way that coded data does, you can verify the summary against the original without too much effort, and the task plays to the model's genuine language capabilities. Use it here with reasonable confidence.
Prompt for structured summary:
You are analysing a qualitative interview transcript. Summarise this transcript in the following structure: 1. Main themes discussed (3-5 bullet points) 2. Key quotes (3-5 direct quotes from the participant, preserved verbatim) 3. Participant's overall stance or position on the research topic 4. Any unusual or unexpected perspectives not captured in the bullet points Transcript: [paste transcript here]
Prompt for extracting specific content:
Read the following interview transcript and extract everything the participant says about [specific topic]. Quote directly where possible. Note if they gave conflicting or nuanced answers on this topic. Transcript: [paste transcript here]
Prompt for executive summary (applied research):
Summarise the following interview transcript in 200 words or fewer. Focus on actionable findings relevant to [specific decision or question]. Prioritise specifics over generalities. Transcript: [paste transcript here]
For large-scale summarisation across many transcripts, tools like Skimle handle batches automatically with consistent output structure — but for a small set, the manual prompt approach works.
Prompts that work: initial code generation
Initial code generation — identifying a preliminary list of codes from a single transcript or a small set — is a reasonable use of ChatGPT, with an important caveat. Treat the output as a starting point for your own analysis, not as the analysis itself. The model will produce codes that look plausible and are often useful, but it will also collapse nuance, miss emotionally loaded content that a human reader would notice, and occasionally generate codes that reflect its general knowledge of the domain rather than what the specific participant actually said. Review everything against the transcript before adding it to your codebook.
Prompt for inductive coding:
Read the following interview transcript and generate an inductive code list. For each code: - Give the code a short label (2-5 words) - Write a one-sentence definition - Quote one example passage from the transcript that illustrates the code Generate codes at a granular level first; do not immediately collapse into broad themes. Transcript: [paste transcript here]
Prompt for deductive coding against an existing framework:
I am coding interview data against the following theoretical framework. Apply these codes to the transcript below. For each code, identify all relevant passages and quote them directly. Codes: [list your codes and their definitions] Transcript: [paste transcript here]
Prompt for checking your codes against a transcript:
I have developed the following codebook from earlier interviews. Read the new transcript below and: 1. Identify passages that fit existing codes (name the code and quote the passage) 2. Identify any passages that do not fit existing codes and suggest new code labels for them Codebook: [paste your codebook] New transcript: [paste transcript]
The third prompt is particularly useful for iterative coding — you are not starting from scratch with each transcript but using ChatGPT to check new data against what you have already identified.
Prompts that work: theme development
Theme development from a code list is harder for ChatGPT than initial coding, and this is where you should be most sceptical of the output. Theme development requires synthetic judgement about relationships between codes — which codes belong together, what underlying pattern connects them, what analytical claim the grouping makes. This is precisely the kind of interpretive work where the model tends to produce output that looks right but reflects its training data rather than your specific dataset. The themes it generates will often map onto well-known frameworks in your domain rather than emerging from the particulars of what your participants said.
Use these prompts to generate a starting structure that you then interrogate critically, not as a substitute for that interrogation.
Prompt for clustering codes into themes:
I have the following codes from a thematic analysis of qualitative interview data. Group these codes into themes. For each theme: - Give it a descriptive name that makes an analytical claim (not just a topic label) - List the codes that belong to it - Write a 2-3 sentence description of what the theme means Aim for 3-6 themes. Do not create a theme with only one code unless there is a strong reason. Codes: [paste code list]
Prompt for writing theme descriptions:
I am writing up a thematic analysis. For the theme below, write a 150-200 word description suitable for a results section. The description should: - State the analytical claim the theme makes (not just its topic) - Reference the codes it contains - Suggest what kind of illustrative quotes would support it Theme name: [name] Component codes: [list codes] Example passages: [paste 2-3 example quotes]
Prompts that work: question design
ChatGPT is genuinely useful for refining interview questions. The risks are lower here — the output guides data collection rather than constituting analysis — and the model's language capabilities are directly applicable.
Prompt for generating interview guide questions:
I am designing a semi-structured interview guide for a study on [topic]. The research question is: [your research question]. My participants are [description of participants]. Generate 8-12 interview questions that: - Are open-ended (no yes/no answers) - Are ordered from general to specific - Include at least two probing questions for areas where participants may give surface-level initial answers - Avoid leading the participant toward a particular answer
Prompt for improving existing questions:
Review these interview questions and suggest improvements. For each question, note any leading language, ambiguity, or structure problems, and suggest a revised version. Questions: [paste your questions]
For the full methodology behind interview guide design, see how to write a perfect interview guide.
Where ChatGPT falls short for rigorous work
The architectural problems described above have specific practical consequences. Worth naming them explicitly, because they are not obvious until you have encountered them in your own data.
Quote hallucination. This is the most serious issue and the one researchers most frequently underestimate. ChatGPT sometimes generates quotes that sound like something the participant said but are not verbatim from the transcript. The quote will be plausible in tone and content — that is what makes it dangerous. Verify every quote the model produces against the original transcript before using it in your analysis or write-up. This is not optional.
Consistency drift. In a single session, ChatGPT will apply codes fairly consistently. Across sessions — which is how most multi-transcript projects work — it will not, because there is no persistent state and no fixed interpretation of ambiguous codes. The same passage coded as "resistance to change" in one session might be coded as "scepticism about leadership" in another. You can manage this by pasting your codebook into every session, but the burden of maintaining consistency falls entirely on you.
Domain knowledge contamination. Because the model has been trained on large amounts of academic and professional writing, it "knows" what themes typically appear in research on workplace culture, healthcare experiences, educational outcomes, and most other common domains. It will produce those familiar themes whether or not your participants actually expressed them. This is the subtler and more intellectually dishonest failure mode — the output looks valid precisely because it matches what the literature says, regardless of your specific data.
Methodology superficiality. Specifying "use Braun and Clarke's reflexive thematic analysis" in your prompt will not produce a meaningfully different output from specifying nothing at all. The model will generate something that looks like thematic analysis because that is what most qualitative analysis writing looks like. The epistemological commitments of different methodological traditions are not operationalised in prompt-based tools.
For the full argument about why general-purpose tools are structurally unsuited to serious qualitative work, see AI text analysis tools: a comparison. For why RAG-based retrieval approaches do not solve the problem either, see why RAG fails for structured qualitative analysis.
Using ChatGPT alongside a dedicated analysis tool
The most productive use of ChatGPT in a qualitative workflow is not as a replacement for systematic analysis but as a supplement for specific sub-tasks:
- Before the main analysis: use ChatGPT to generate preliminary summaries and a provisional code list from your first few transcripts. This helps you develop your codebook faster.
- During analysis: use it to draft theme descriptions from your finalised code clusters, or to check specific transcripts against your codebook.
- After analysis: use it to draft section of your write-up from your theme descriptions and selected quotes. The guide on writing up a thematic analysis covers the structure you should be working toward.
For the main analysis — systematic coding across your full corpus, consistent theme application, and a traceable link from every insight back to its source quotes — a dedicated tool handles this more reliably. Skimle's approach is to structure the analysis as a defined process with consistent output, so the AI's contribution is auditable and the human researcher's interpretive judgements are clearly preserved. See how Skimle handles the analysis process for detail.
A note on academic use and attribution
If you use ChatGPT in your qualitative analysis, you need to document it. The minimum required in a methods section is:
- Which tool you used and in what role (e.g., initial code generation, transcript summarisation)
- How you validated and refined the AI's output through your own interpretive work
- A clear statement that analytical judgements were made by the human researcher
More journals are accepting AI-assisted analysis when it is documented transparently. What reviewers push back on is undisclosed use or claims that the AI "did the analysis" without human interpretive oversight. See the AI in qualitative research guide for academics for a fuller account of documentation requirements across different publishing contexts.
Want AI-assisted qualitative analysis that is systematic, traceable, and ready to document? Try Skimle for free and see how structured AI analysis produces a quote-linked theme hierarchy you can write directly from.
Related reading:
- Can ChatGPT analyse qualitative data?
- AI text analysis tools: a comparison
- How to use AI in qualitative research: a guide for academic researchers
About the authors
Henri Schildt is a Professor of Strategy at Aalto University School of Business and co-founder of Skimle. He has published over a dozen peer-reviewed articles using qualitative methods, including work in Academy of Management Journal, Organisation Science, and Strategic Management Journal. His research focuses on organisational strategy, innovation, and qualitative methodology. Google Scholar profile
Olli Salo is a former Partner at McKinsey & Company where he spent 18 years helping clients understand the markets and themselves, develop winning strategies and improve their operating models. He has done over 1000 client interviews and published over 10 articles on McKinsey.com and beyond. LinkedIn profile